Multimodality radiomics prediction of radiotherapy-induced the early proctitis and cystitis in rectal cancer patients: A machine learning study.


This study aims at predicting radiotherapy-induced rectal and bladder toxicity using computed tomography (CT) and magnetic resonance imaging (MRI) radiomics features in combination with clinical and dosimetric features in rectal cancer patients.
Methods:
Sixty-three patients of local advanced rectal cancer who received radiotherapy were included, and the toxicities including proctitis and cystitis were scored for rectum and bladder, respectively. The pretreatment planning CT and MR-T2W-weighted image radiomics features were extracted from the rectum and bladder regions of interest, and Lasso, MRMR, Chi2, Anova, RFE, and Selectpercentile were used for feature selection. Machine learning algorithms including K nearest neighbor (KNN), support vector machine (SVM), logistic regression (LR), decision tree (DT), random forest (RF), naive bayes (NB), gradient boosting (XGB), and  linear discriminant analysis (LDA) classifiers were used for predictive modeling. The effect of Laplacian of Gaussian (LoG) filter was investigated with sigma from 0.5 to 2. The models were compared in terms of prediction accuracy, precision, sensitivity, and specificity.
Results:
The highest predictive performance for the pre-MRI T2W model had accuracy: 90.38/96.92%, precision: 90.0/97.14%, sensitivity: 93.33/96.0%, and specificity: 88.09/97.14%, with original image/LoG filter (σ=0.5-1.5) based on LDA/DT-RF classifiers for the proctitis/cystitis, respectively. Furthermore, for the CT scan, the accuracy: 90.38/96.92%, precision: 88.14/97.14%, sensitivity: 93.33/96.0%, and specificity: 88.09/97.14% had the highest value based on XGB/DT-XGB classifiers for the proctitis/cystitis with LoG filter (σ=2)/LoG filter (σ=0.5-2), respectively. MRMR/RFE-Chi2 feature selection methods had the best function of proctitis/cystitis, respectively, for the pre-MRI T2W model, and also MRMR/MRMR-Lasso had the highest model performance for the CT.
Conclusion:
Radiomics features derived from pretreatment CT and MR images could predict radiation-induced proctitis and cystitis. We observed that LDA/DT/RF/XGB classifiers with MRMR/RFE/Chi2/Lasso feature selection algorithms with LoG filter had a good predictive performance.&#xD.